Deep learning techniques for the fully automated detection and segmentation of brain MRI

Over the past decade, auto-segmentation for tumors has drawn a lot of attention due to its significant impact on cancer treatment. Auto-segmentation architectures have a significant role in alleviating the enormous workload on the medical staff. This has motivated us to explore the latest solutions in auto-segmentation to use it in auto-segmentation. It works on automatically contouring tumors to make radiology treatment more attainable since manual contouring is repetitive and subjective to human error. Auto-segmentation usually strives to achieve high accuracy to reduce the time the radiologists take to contour the tumor. Saving time is critical as instead of contouring all the tumors, the radiologist can spend the time editing on the segmented tumor thus more patients can be diagnosed in less amount of time. There have been a lot of auto-segmentation architectures created for general purposes like the Segnet which is sometimes used in medical segmentation, but such architectures fail to achieve high accuracy especially in the details of the tumor. The U-Net is an auto-segmentation architecture specifically created for auto-segmentation on medical images like MRI and CT. The U-Net architecture can achieve high accuracy of segmentation with fewer amounts of data. We improved U-Net performance by using residual blocks on each layer of the architecture itself usually referred to as Res-U-Net. Our final proposed fine-tuned Res-U-Net model has achieved 97.10% on the used data which was the best of our 3 proposed fine-tuned models. The used data was Low-grade gliomas (LGGS) brain tumor dataset.